Firstly, a filesystem large enough to trigger inode32 rotor behavior is needed. This can be achieved on a small local disk by creating a large sparse file, writing a new filesystem onto it and mounting it in loopback mode.

> dd if=/dev/zero of=loop bs=4K seek=1G count=1 > /dev/null 2>&1
> ls -sh loop
4.1T loop
> mkfs.xfs -d file=loop loop
meta-data=loop                   isize=256    agcount=32, agsize=33554432 blks
         =                       sectsz=512   attr=0
data     =                       bsize=4096   blocks=1073741824, imaxpct=25
         =                       sunit=0      swidth=0 blks, unwritten=1
naming   =version 2              bsize=4096  
log      =internal log           bsize=4096   blocks=32768, version=1
         =                       sectsz=512   sunit=0 blks
realtime =none                   extsz=4096   blocks=0, rtextents=0
> xfs_db -f -c "sb 0" -c "p" ./loop | egrep 'agcount|inopblog|agblklog'
agcount = 32
inopblog = 4
agblklog = 25
> sudo mkdir /mnt/loop
> sudo mount -o loop loop /mnt/loop
> df -h /mnt/loop
Filesystem            Size  Used Avail Use% Mounted on
/home/sjv/loop        4.0T  528K  4.0T   1% /mnt/loop
> sudo chmod 777 /mnt/loop
> cd /mnt/loop

Next, ten files are created in a directory on the new filesystem and their inode numbers and data extent allocations are examined:

> dir=a
> mkdir $dir
> for file in `seq 0 9`; do xfs_mkfile 10m $dir/$file; done
> ls -is $dir/*
> for file in `seq 0 9`; do
> ag=`xfs_bmap -v $dir/$file | tail -1 | awk '{print $4}'`
> echo $dir/$file is in AG $ag
> done

Repeat step 2 with a different rotorstep value:

> sudo sysctl -w fs.xfs.rotorstep=3
> dir=b